Research brief · Lakebed by Theo · 2026-05-29

Lakebed is not “a cloud provider” yet. It is an agent-native mini-app runtime with hosted deploys.

The useful mental model: Lakebed is trying to make the smallest full-stack TypeScript app into one self-contained “capsule” that an AI coding agent can understand, build, inspect, and deploy without assembling Next.js + DB + auth + hosting + logs from scratch.

Use it for

throwaway internal tools, agent-built UIs, demos, small stateful apps, webhook experiments.

Do not use it for

production systems, complex npm dependencies, file storage, long-lived critical data, serious auth/compliance.

Payment model

No public pricing or billing surfaced. Current public flow is anonymous deploy → GitHub-claimed deploy.

Bottom line

Recommendation: spend a 90-minute spike on Lakebed, but treat it as an exploratory agent-app substrate, not infrastructure to migrate to. The immediate value for Connor is learning its capsule contract and testing whether Hermes/Codex can reliably generate useful single-purpose apps with fewer moving parts than the normal web stack.

Value signal

High if the goal is “let agents create small tools quickly and inspect state/logs cheaply.” This aligns with Hermes / Personal Orchestrator style workflows.

Maturity

Alpha / early prototype. Public docs explicitly call it “alpha-stage”; npm README calls it an “early public prototype.”

Production readiness

Low today. No arbitrary npm imports, no Node built-ins inside capsules, no file storage, local state resets, limited schema types.

What Lakebed is

Lakebed describes itself as “an alpha-stage agent-native CLI and runtime for building small full-stack TypeScript apps called capsules.” A capsule is the whole app: server contract, Preact client, state, auth, logs, and deploy URL.

The “agent-native” part matters. The docs tell agents to treat the capsule directory as the whole app: write server/index.ts, write client/index.tsx, run the CLI, inspect DB/logs, and deploy. That is much narrower and more machine-manageable than the usual sprawl of package config, framework routes, ORM migrations, auth provider config, hosting, log drains, and database dashboards.

server/index.ts      schema, queries, mutations, HTTP endpoints
client/index.tsx     Preact UI using Lakebed hooks
shared/              pure TypeScript shared by server and client
.env.lakebed.server  optional server-only env values

Feature breakdown

AreaWhat Lakebed gives youWhat is missing / constrained
App modelOne capsule = source + server API + client UI + state + auth + logs + deploy URL.No general framework surface; v0 expects one server entry and one client entry.
Servercapsule() with schema, queries, mutations, and external HTTP endpoints/webhooks.Server modules import only lakebed/server and pure relative files; no Node built-ins in capsule modules.
DatabaseTyped table helpers; rows get id, createdAt, updatedAt; methods include where, orderBy, limit, all, get, insert, update, delete.V0 field helpers are basically string(), boolean(), and defaults. Not a full Postgres app model for capsule authors.
ClientPreact app with useQuery, useMutation, useAuth, routing, links, Google sign-in UI.No arbitrary npm imports; Tailwind classes in JSX but no CSS pipeline/config.
AuthGuest auth by default; Google sign-in built in. Dev can switch guest identity with npx lakebed auth as alice or per-tab query params.Not a full auth platform. Treat built-in auth as useful for demos / simple apps until proven otherwise.
Env / secrets.env.lakebed.server is available to server handlers through ctx.env; not exposed to client or embedded in anonymous artifacts.Hosted env sync requires a claimed deploy. Sync replaces hosted env with file contents.
InspectionCLI can inspect DB, dumps, logs locally and hosted: npx lakebed db dump, logs, inspect.Public inspection must be explicitly enabled. Private hosted inspection relies on claim token in .lakebed/deploy.json.
Deploynpx lakebed deploy publishes anonymous deploys; claimed deploys can use env and outbound server-side fetch; claimed apps can reserve *.lakebed.app subdomains.Anonymous deploys disable outbound server-side fetch and env. Unclaimed anonymous deploys expire/clean up in runner defaults.
Self-hosting / runnerPackage includes anonymous-server; README documents Postgres via DATABASE_URL and Railway-oriented service split.Public repo appears unavailable/private; self-hosting is possible from npm package but not yet a polished operator experience.

Cloud provider and payment model

Hosted Lakebed

The CLI defaults to https://api.lakebed.dev; dashboard is https://dashboard.lakebed.dev/deploys; app base domain is lakebed.app. Deploy flow is anonymous first, then claim via GitHub OAuth when you need server env or outbound fetch.

free anonymous deploy path existsresource-limitedalpha

Payment model

I found no public pricing page, billing page, paid tier, Stripe checkout, or usage price list. The public product model currently looks like “try/deploy anonymously; claim with GitHub for non-expiring deploys + env/fetch.” That does not mean it will stay free; it means pricing is not publicly specified.

no public pricingGitHub claim

Known limits from docs/source

Self-hosting / cloud provider angle

Lakebed can run a deploy API/runner yourself. The README documents production env such as PUBLIC_ROOT_URL, LAKEBED_DASHBOARD_ROOT_URL, LAKEBED_APP_BASE_DOMAIN, DATABASE_URL, GitHub OAuth env, session secret, and server env encryption secret. It specifically references a staged Railway split: all-in-one, api-dashboard, and capsule-runner services.

So the cloud-provider story is: hosted Lakebed runs Theo’s managed API/dashboard/runner, while the package contains enough runner/API machinery to run it on your own infrastructure, especially Railway/Postgres. It is not yet a general “bring any cloud provider and click deploy” product.

How this could be valuable for Connor

1. Agent-generated internal tools

Lakebed’s strongest fit is “Hermes, make me a tiny app for this workflow.” Examples: Open Tabs triage UI, proposal review queue, GBrain query workbench, audio-memory approval panel, personal finance/admin checklist. The capsule boundary is simple enough for agents to reason over.

2. Disposable stateful previews

For experiments that need state/auth/logs but not a full stack, Lakebed could be faster than bootstrapping Next/Prisma/Supabase/Vercel. Anonymous deploys are especially useful for sharing a prototype quickly.

3. Runtime inspection as a design primitive

The built-in DB/log inspection is aligned with how you evaluate agents: prove done, inspect state, avoid guessing. A capsule can become a small, inspectable artifact rather than a pile of framework files.

4. A substrate to study

Even if you do not adopt Lakebed, its contract is worth studying: “agent-native app runtime” is exactly the kind of bounded surface Hermes could target for personal software generation.

The key caveat: Lakebed currently compresses the app stack by removing degrees of freedom. That is good for agents and demos; bad for complex production requirements. The value is not “replace Cloop infra”; it is “create a fast lane for small, agent-authored software.”

Best next steps

Step 1 — 20-minute local smoke test

mkdir -p ~/experiments/lakebed
cd ~/experiments/lakebed
npx lakebed new open-tabs-mini --template todo --no-git
cd open-tabs-mini
npx lakebed dev

In another terminal:

npx lakebed auth as connor
npx lakebed db dump --port 3000
npx lakebed logs --port 3000

I verified this environment can create and build the starter capsule with Node v22.22.0 and npm 10.9.4. The anonymous build produced a Lakebed artifact successfully.

Step 2 — build a Connor-relevant micro-app

Do not start with a generic todo. Build one useful capsule:

Step 3 — deploy anonymously, then decide whether to claim

npx lakebed deploy
# If you need env or server-side fetch:
npx lakebed claim
npx lakebed deploy

Claim when the capsule needs secrets, external API calls, or permanence. If it is just a demo, anonymous deploy is enough.

Step 4 — evaluate with a hard rubric

QuestionPass condition
Can Hermes/Codex edit the capsule without confusion?Agent can implement a feature and pass manual inspection in under 20 minutes.
Is state/log inspection useful?db dump and logs explain behavior faster than reading code.
Can it call your real systems?Claimed deploy with env + server fetch can hit a harmless webhook or test endpoint.
Is the restricted model painful?If you immediately need npm packages, file storage, custom CSS, or complex data types, Lakebed is not ready for that use case.

Risks and open questions

Nerd Snipe transcript note

The most recent Nerd Snipe episode in the Substack RSS feed is “Google is Not a Serious Company” from 2026-05-28. Its description says: “Oh and apparently Theo's building his own cloud.” I attempted to pull the YouTube transcript for the embedded video, but YouTube caption access was blocked from this cloud IP and the local YouTube-memory fallback did not recover usable subtitles. I therefore used the episode metadata as a directional clue, not as a source of detailed claims.

Sources